Software Installation and Management

SAInt computation capabilities are designed to take full advantage of modern CPUs. SAInt performance is independent of the availability of a graphic card (GPU) for both simulation and optimization problems. For SAInt’s typical optimization workloads (LP/MIP using Gurobi), GPUs do not accelerate solves; performance primarily depends on the CPU and memory. Furthermore, SAInt currently runs Gurobi without enabling the "Primal-Dual Hybrid Gradient" algorithm on GPU (PDHG-on-GPU), as this is still considered a beta feature for NVIDIA GPUs in release 13.0 of Gurobi.

More cores can improve performance, especially for models requiring extensive parallel processing. For problems that solve faster with linear programming (LP) or mixed-integer programming (MIP) near the root node, higher clock speeds are preferable.

We have observed that on machines with more than 32 cores, there is a little-to-no benefit for a single SAInt solve, where clock frequency and RAM play a more important role. On the other hand, if the machine is used for multiple concurrent runs of SAInt, a higher number of cores could help.

In optimization problems, Gurobi has a soft limit of ~32 threads, since higher thread counts often don’t improve performance. This is the default automatic option by Gurobi (i.e., Threads=0). On machines with >32 logical processors, under default settings, Gurobi may use only the number of physical cores unless Threads is explicitly set higher (e.g., 24 physical cores with 2 threads/core then set Threads=48 to use all logical threads).

More RAM is beneficial for handling large and complex models. Moving from 16 GB to 32 GB of RAM can yield major speedups if 16 GB causes paging/memory pressure. RAM should be provisioned in proportion to the model size and complexity. For many typical single-run cases, performance improvements taper off once ~32 GB is sufficient to avoid memory pressure/paging; beyond that point, additional RAM usually yields limited speedup unless the instance is large enough to materially increase working-set memory needs. When multiple SAInt runs are active on one machine, the provision of more RAM is very beneficial since each process will consume substantial memory. Splitting the RAM into more memory cards to leverage more memory channels, could help enhance performance.

If possible, prefer ECC (Error Correcting Code) RAM over non-ECC RAM. ECC RAM’s main benefit over "normal" (non-ECC) RAM is data integrity: it can detect and often correct bit flips that would otherwise silently corrupt data or crash a program.

A machine with more cores and a slightly slower clock speed may outperform a machine with fewer cores but a faster clock speed (or vice-versa), depending on the nature of the optimization tasks. Benchmarking specific models is crucial to determine the ideal configuration.

If planning to run multiple instances in parallel, ensure sufficient cores and memory to support concurrent processing without resource contention. Limit the number of threads per process if necessary to balance the computational load.

CEM is RAM-intensive, and it is recommended to have at least 128 GB of RAM. We strongly recommend reducing the time horizon and / or the problem size to fit within hardware constraints, where possible.

It is recommended to equip the workstation or server used for simulation and optimization problems with an adequate and efficient cooling system, especially if the machine needs to operate for long periods. Lower operating temperatures help in maintaining higher sustained all-clock frequencies and avoid thermal throttling.